Pax (00:00)
Hello everyone, I'm Paxton Gray, CEO of 97th Floor, and this is the campaign.

Thank you for joining us today for another episode of the campaign where we talk with marketing leaders about better knowing your audience, innovating beyond best practice, and converting visitors into customers. The campaign is produced by 97th Floor.

A world-class digital marketing agency designed to build organic and paid channel strategies for mid-level and enterprise organizations. You can find past episodes of the campaign on iTunes, YouTube, Spotify, and at 97thfloor.com. Today's guest is Kevin Indig. Kevin has led organic growth at Atlassian, G2, and Shopify, and now advises companies like Meta, Ramp, and Upwork on building cost-efficient acquisition channels. In this episode, Kevin is going to share a ton of his research. I have

bunch of questions about how to optimize for AI search and how brands should be rethinking their plays from SEO now into AEO slash GEO. Let's get into it.

Pax (00:59)
Kevin, thank you so much for joining the show today. It's a pleasure to have you on again. Looking forward to talking with you.

Kevin Indig (01:03)
Man, it's

Man, it's always nice to be here.

Pax (01:06)
So today we're gonna talk all about AI search. And I think a a good framework would help set the stage here. when it comes to thinking about AI

visibility specifically as a pipeline. you talk about, you know, retrieved, cited, and trusted as the main stages of that pipeline. So tell us about that.

Framework and then where do you see brands kind of dropping the ball within those stages?

Kevin Indig (01:32)
The framework came from me thinking a bit about the differences between pre-AI SEO and SEO or AEO today, or whatever you want to call it, right? But that's where it comes from. And so the pipeline, as you said, is retrieved, cited, trusted. it's different than crawled, rendered, indexed, and ranked, for a good reason. And that is because we are in a new world.

And so the first step retrieved is critical because I I noticed that a lot of people have the mental model in their heads that you know you're being crawled and then cited and then mentioned in AI answers. And that is not completely accurate. they go through different steps. So they come, they can come to information or to your brand.

Either through training data or they do live web retrieval. So that is what we most commonly understand under RAG, retrieval augmented generation. And it is a crucial step to reduce hallucination. And so when LLMs realize, hey, I don't have all the information I need in my training data, they go out to basically perform a web search or several web searches. They find URLs in those web searches, they crawl these URLs, they they

Retrieve the information from these URLs, but that doesn't automatically mean that you're cited. This is a really important misunderstanding that I realized. And that is just because an LLM crawled your content, it does not mean that it's automatically cited. It just means that it is in a pool of sources to be considered for citation. A tiny chunk of that pool will actually be cited, which means it is linked to as a source in the AI answer.

And then you have a chance to show up as a brand mention in the AI answer based on what the citations are saying. And then we have the final step, which is trust. And trust is so critical because I saw over and over again in my studies that trust is really the most important ingredient when it comes to AI period. Not even just talking about AI search, but AI period. And the reason is that people select.

Brands that get recommended based on trust. So, quick step back: I ran this user behavior study where we recruited a panel of 50-ish US adults. We gave them tasks in search. So they were supposed to shop for high-priced items like washing machines or laptops. And we observed their user behavior and their clicks and what they what they say about their experience.

We found is that when users get a shortlist in AI. So for example, in AI mode, they ask for one of the best laptops. For me specifically, they get maybe five to ten products. That comes as a short list. 75% of the time, users will pick the first result on that short list, right? They pick the number one recommendation. The only exception to that is when they see a brand on that shortlist that they trust.

In that case, it doesn't matter where the brand is. If they trust it, they will always prefer it. And that's why it is retrieved, cited, and trusted, because all of the hassle going through all the steps of the pipeline will not help you if you're not at least a somewhat trusted brand.

Pax (04:47)
Okay, so when you say trust, you're not referring to trusted by the algorithm or the LLM, you're saying trusted by users. have you noticed a relationship between cited brands and brand mentions? brand mentions, as you said, was based off of the information in the cited content. And there's the play, obviously, of

citing yourself can help, if your own citation mentions yourself, you can be mentioned as a solution. Taking that out of the equation, have you noticed any correlation between brands that get cited more often with brands that get mentioned more often?

Kevin Indig (05:27)
Is a correlation, but it's not as strong as people think. a lot of times, brands that are the citation for an answer that are sort of say the source of an answer, they're not automatically mentioned, but at the same time, mentions matter so much more, right? So, people often ask me what is the right metric and what should we optimize for? At the end of the day, it is mentions, ideally, mentions in the right context, right? Ideally mentions as

First in a short list or better than a competition or in a positive light, But mentions are really what users see as a sort of recommendation at the end of the day, and that influences their purchase behavior the most. What I've also seen is that the ratio of being mentioned and being cited varies by LLM. So for example, Gemini will

have a very low citation count but mentions a lot of brands. Whereas ChatGPT cites a lot of sources but doesn't mention that many brands in like a relative context. Right. So it's also critical that we think about that ghost citation problem in the context of different LMs.

Pax (06:30)
back to your initial finding Have you noticed behaviors varying between different models or platforms? I've heard that concept a lot in the context of Chat GPT.

going and doing manual searches and then pulling information back from that. Have you seen the same with like AI mode, for example? Or is Google's index like it's kind of one and the same. And so therefore you're not getting live searching from like an AI mode answer.

Kevin Indig (06:59)
We have reason to believe that Google uses separate indices for AI search versus quote unquote classic search. And then of course there's a whole kind of training data aspect that also comes on top of that. And it does seem that AI mode and Gemini use different versions of the Gemini model. But he made an important point there, which is that there's not just a difference between LMs but also models. So when OpenAI rolled out ChatGPT 5.5, we saw a distinct jump in.

The number of fan out queries that ChatGPT will search for compared to 5.3. So fan out queries is essentially the searches that LLMs conduct when they use web search. And they use web search quite a lot. and it has profound impact because we know that search rank is still very, very important when it comes to AI visibility. but the question is what search rank? And so

It's the query fan out search rank, right? That is the the the search rank that matters. but a lot of times these fan out queries don't have search volume, right? So one distinct difference from classic SEO to new AI SEO, whatever you want call it, is that we have to optimize more for keywords that might not have search volume, at least not from humans, but from bots.

Pax (08:11)
Mm, interesting. So how do you then identify those queries other than just, you know, running these prompts and seeing what the query fan out generates? Is that really like all you have to do? Or and then if you're looking at different models, you know, you're looking at potentially like an unending research process or data gathering, like what's your best practice there?

Kevin Indig (08:34)
It is a never ending research process. there's there's a lot of data, right? because to your point, it's not just the models, it's the LLMs, et cetera. So at the end of the day, the best way to go about this today is to use tools.

you look at the fanal queries of each model and you basically add it up, right? Because if you care about Gemini and you care about ChatGPT, or you care about Copilot and Perplexity, at the end of the day, you then need to be visible and rank for all of the fanal queries that these models are going after. Now

One last thing about the models, right? I also see too many aggregate scores. I see too many teams chasing all the models and then tracking a model that maybe they don't even care about. Or it better said that their customers don't care about. So you want to be very intentional about not just which model you monitor, where you track your prompts, but also where you spend your time or what you spend your time.

Optimizing for because the reality is there's a lot of models right now, there's a lot of platforms right now, and optimizing for all of them with the right level of dedication intensity is just very feasible for anyone.

Pax (09:46)
Have you found a good way to get a good status quo of like here's what

these LMs think about you now so that brands can start to strategize on how to alter that.

Kevin Indig (09:57)
So Dan Petrovic coined this term of selection bias. And essentially, you know, if you think about a single prompt, it's probably hundreds of prompts or a thousands you should be thinking about. But if you think about a single prompt, say a category prompt, something like, what is the best sneaker for running? You type that in once, you get a you get a list of brands, right?

take a set of prompts and run them.

At least I would say 20 times in a row consecutively, and then see what shows up consistently, right? What brands, what attributes, what structures of the answer. and then you get a good understanding of your bias and how stable you show up in that specific model. Now, we said this before, we probably have to do this for every model, so it adds up a little bit, but that's the most reliable way right now to understand what a model truly thinks about us.

And how consistently we actually show up in a category.

Pax (10:51)
I love that. you introduce a really great point we talk about in AI search this specificity around the prompts. And you've written about like prompts are often, three to ten times more words than the traditional like search terms that we're used to.

It represents this fracturing of attention where you have these very specific prompts. as you kind of alluded to, you also have the historic context that we're not used to at all with Google search, which is.

Where has this user been? Who is this user? What do I already know about them? And you know, person A with a different history will get a different answer than person B with for the same prompt, effectively. something in my gut tells me you know, I love the measurement and I love wanting to know everything that's happening, but there's a certain point where it gets so large.

That it seems almost impossible to wrap your arms around it. And it seems like brands will be better served by, in some ways, dropping some of the data and focusing a little bit more on audience and just say, just focus on this niche. Don't worry about how many people are searching. if you just serve this audience, you're gonna be great in a year or two from now.

I'd love to hear your kind of reaction given that you are so deep in the data.

Kevin Indig (12:09)
Totally with you. You know, I think prompt tracking is closer to polling than to rank tracking. we've gotten so used to rank tracking and to simplifying things and like one search engine, it's like a very simple model, position up, down, zero-sum game, 10 ranks or 10 positions. and that is out the window now, right? It's much more like an occasional read on where things are.

I mean, I can probably riff on an hour for an hour about like all the different ways that prompt tracking needs to be set up to be closer to reality. And it's still not going to be perfect, right? It's an approximation because we are now in a world where basically everybody has a unique experience in AI chatbots, right? But the fact that prompts are getting so long, the likelihood that somebody asks the same prompt is very close to zero in most cases. So

I love your push towards audience because I truly think that as marketers we have forgotten how to talk to people because performance marketing has gotten so good that we just basically need to create, you know, target profiles, creative messaging, and boom, customers come in, right? what we need to do not just to understand our audience and how to appeal to them better, but also to

understand how they perceive us is to go back to classic brand surveys, you know, polling, focus groups, interviews, all these kinds of things. The good news is that AI makes the evaluation and measuring so much easier, right? AI is great at turning unstructured text into structured text. So this is a good time, right, to go back to that. But we need to relearn and and rebuild and atrophied muscle.

Pax (13:43)
Mm-hmm. Yeah, I love that. I love that direction. we for years, you know, marketers have been going off of keyword volume data as the inspiration for their content. And if you think about it, it's like everyone's been using this same set of data to produce their content. And so what hope does anybody have of actually standing out and doing something new? And there's also a huge set of things that

People could use that they're not asking for because they don't know it exists yet. And so it will never show up in any kind of volume or prompt tracking data because they haven't known to search for it yet. shift gears a little bit. it's been said a lot in the industry that uniqueness is a very big factor. Adding something to the conversation is really important to get cited. your research found that original data is one of the strongest single predictors.

of page originality and that pages with 15 or more unique data points score way higher on information gain, quote unquote, than pages with one or fewer. tell me more about this information gain metric and how brand should think about that.

Kevin Indig (14:48)
So information gain as a as a as a principle, as a paper, even, has been around for a long, long time but it's gotten a new reputation even before AI entered the scene, because SEOs were realizing, hey, we're all creating more or less the same content. what gives, right? Like how do you stand out here? And then

There is some rumors of Google using or quantifying information gain, which are probably true. But I think a lot of people underestimate how much content there is and how much content search engines and LMs have to sift through to even just get to the top 100 or maybe top 10. information gain essentially describes the

gain in understanding that users have after consuming your content relative to all the other content. Or in other words, what do you have that is relevant, which is critical, that other content does not have right now, a a great way to facilitate that is is primary research. Right. if you're a software company or even as a consumer company, you look at your product usage, you you look at your customers, you look at the market.

What is data that only you can collect that you can turn into a story that is relevant to your audience? companies that publish primary research tend to be cited more often. And especially companies that publish reports, most often benchmarks, but reports.

Or research that addresses real buyer conversations and real buyer questions. So, in other words, sure, you can publish a study about something, but if that something is not relevant to your audience while evaluating your product or being in the market, then it doesn't have that citation boosting effect.

And also, by the way, to like get attention and form relationships with your audience and foster trust. Right. There's a lot to be said about primary research that speaks for you, but in a pure AI search lens, it

you know, like making it tangible to market questions is the best way to go about.

Pax (16:47)
Sure. Especially if you take in the fact that like, do you really want to be manipulating what are ultimately like, yes, the LLM, but do you want to manipulate users and risk, you know, betraying their trust and kind of torching your company's reputation in the long term? so for those brands that say, Okay, great, I want to publish some proprietary data, but you know, I don't have a million users that I can

pull data from. I don't have a giant data set. What would you recommend to these companies who maybe feel like they don't have obvious proprietary data but still want to publish some benchmarks?

Kevin Indig (17:23)
There's a ladder. At the top of the ladder is, as I mentioned, product usage. What can you learn about how your customers engage with your product? one rung lower on that ladder is basically publishing data about the market, right? So you can you can maybe scrape data, you can download data sets and maybe combine them. so what can you put together that is still unique, you know, maybe not published, but it's not coming directly from everyone.

And then lastly, our surveys and polls and market research. and I will say that, that's probably the most accessible one, but it's also the easiest to copy. I think what's key there is not just the data you can get.

But also the conversation you join. So the biggest mistake that I've seen brands make when it comes to data storytelling is that they just look at what they have and then they spin a story around it. Sometimes that goes well. Most of the time it doesn't. The way you want to approach it is you want to see what's going on in the like world at large, but also the world of your customers, meaning your market. And then what can you speak to with data? Right. So if you're

an accounting software company, and there are new regulations coming out that affect your clients, that is definitely worth writing about. Maybe you can get some primary data about how many people would be affected and what the effect would be. Would they lose or make money? And you know, like how is that stand in relation to other countries? It's that storytelling aspect that's critical, but it's only gonna get attention.

If you join a conversation that is relevant to the people you're trying to talk to, and that's what most companies miss.

Pax (19:01)
Well let's let's now talk about some external factors. one of them being third parties citing your brand. So you had one data that showed competitor domains hold 33.5% of all of AI citations for invoicing questions, but only 7% for quote unquote starting a business type questions.

So what what are the practical implications for a brand doing PR off site authority building in AI search?

Kevin Indig (19:30)
So we know that third-party authorities are important because LLMs are forming consensus. So when they give an answer, they try to not just base that answer off of the source that is being mentioned in the prompt, right, or being asked about. They also want to source social networks, entertainment platforms, publishers, affiliates, to form a coherent answer, right? To like really

Give the best answer possible that is as objective as possible. and so what's critical is similar to before, right, where we talked about you don't want to have or just have an aggregate score across all LLMs. it's the same idea here. You don't want to think about these third-party authorities across all of your topics as the same. And otherwise, you end up with these, like, you know, very

unactionable lists where usually you'll have Reddit or YouTube at the top and it's like, yeah, okay, so I guess we need to do Reddit. I guess we need to start a YouTube channel. You probably should, right? But for different reasons. What's what's the true implication, is that you want to think about the importance of third-party authorities by topic. So most companies serve several topics based on their products, right? Topic can also be a category, a product category in the consumer space. But if you think about a topic

back in the software space, right? This data that I wrote about, you know, the the company behind this is present in at least 10 different product categories. You need to have at least 10 different topics, if not more. And they need to segment their third-party analysis by topics. So they have 10 different lists for which third parties are most important to engage with or to be present in and to potentially target with PR campaigns. Cause if you think about it, right, if you if you publish a

primary research or a benchmark report about accounting that might not be as relevant to invoicing, right? So these are separate tracks and you need to develop separate campaigns for them.

Pax (21:23)
you talked about

This like third party sources kind of reinforcing this consensus that these LLMs can have. Historically in SEO, we have run away from duplicate content like the plague. It's like avoid it at all costs. You do not want this page appearing anywhere else. Otherwise, you risk penalty, you know, from the old, old days. Do you think that there's maybe an opportunity for that to

be in existence. And I don't mean to game the algorithm, but I al I do mean in the sense that was always, I think, a little misguided because one article living on your site, why shouldn't that live in LinkedIn pulse? And a totally different audience would see that. So do you think that maybe we should re explore the concept of like, hey, take that same stuff and put it multiple places. That would reinforce

This consensus for LLMs and thus more likely become truth, quote unquote, in the eyes of these LLMs.

Kevin Indig (22:24)
100%. I think in this AI search world, we need to pay much less attention to cannibalization, even on our own site. LMs are incredibly good at parsing out nuances and content. They care much more about chunks, etc. But they also, as you said, care about authorities. And so yes, we absolutely should repurpose across different platforms. and I've seen experiments where people even

Copy the like one to one article to LinkedIn and then see really good results. So yes, that works. and I would maybe add like a lens of audience building to it, right? Like there's again, there's many benefits of doing those kind of things. But the reality is that it's not just LLMs that are caring about all these platforms and consensus, also users, right? So if you're a user and you read a really good review on Trustpilot and you go to the website and it has, you know.

Great product shots and they go Reddit and people say this is the worst product I've ever bought. probably a problem for you, right? So, yes, it needs that 360 surround sound type of strategy where you repurpose your content and maybe repost across many different platforms. and you much rather want to get your content out there than fearing duplicate content, especially when it's on different platforms.

Pax (23:34)
speaking of LinkedIn, you you talked about it as a like fast lane. and like brands are entering AI answers within weeks of consistent named author publishing. Do you think that's because LinkedIn is a little undermodeled right now and they're catching up? Or do you think there's something unique about the structure of LinkedIn, like timestamps or connection to professional identity or anything like that, that is making it such a fast lane?

Kevin Indig (23:59)
Connection to identity definitely matters, but there's also a lot of content on LinkedIn, right? they have very strong antibodies against scraping and automation, all that kind of stuff. That being said, I'm very, very unhappy with all the LinkedIn AI comments. I just want to say that for a moment here, that's it's I think it's poison for the platform. But

LLMs care about content that's easy to read and understand, right? So Reddit and LinkedIn, much, much easier to parse because it's text and YouTube. They rely on YouTube video, sorry, transcripts, but still, right, there is that extra hop, that extra context. So there's it, there's the identity part, there's the textual part. LinkedIn is very well optimized from an SEO perspective as well, right? It's one of the largest sites.

And then it's also a question of topical relevance, right? So when it comes to consumer topics, you probably won't find a lot on LinkedIn. So that's where they're less interesting. But when it comes to B2B content, that's where maybe where where Reddit is not that interesting. there's more interesting B2B content on LinkedIn. And I think all these factors together make it a fast lane and make it very interesting.

there's so many benefits you get besides just the AI visibility, right? There's engagement, you talk to your audience, you maybe get some questions that you want to address in your content because people probably ask these questions as prompts on on AI platforms as well. So LinkedIn is a fast lane for several reasons.

Pax (25:19)
You've talked about how a lot of AEO or GEO strategies are just lists of tactics, not really true strategies. what would you say are some of the like hallmarks of an actual AEO strategy versus just a set of tactics?

Kevin Indig (25:34)
A semantic beef with the word strategy, or better said, with how it's most commonly used. and that's it's probably all semantics, right? But the most of the time when people say strategy, they actually mean a plan. They want a plan. The strategy distinctly describes a very specific problem, a reason for why that problem is important, and then a unique approach to solve that problem. you can tell I've read too many strategy books.

But you know, for example, when people say, what is our strategy for SEO? here's our keyword list. That's not a strategy, right? A strategy for SEO or AEO, better said, would be maybe you want to grow your mention rate, maybe you want to change your sentiment, your citation rate, whatever it is, right? It needs to be specific. So where are we today? Where do we want to be? what's the impact on the business? very hard to answer, by the way.

But there needs to be an answer to why does it even matter? Why should we care about this? Otherwise, you get caught up in these like alignment loops. And at the end of the day, you don't get the resources you need. And then lastly, what's your unique approach? And the the keyword here is unique. Most companies do not think about what is our competitive advantage? What is our edge here? What can we do? Do we have access to unique writers? Do we can we scale content a certain way? Do we have crazy authority? You know, like what is our

unique advantage to achieve that goal that other competitors either cannot copy or it's really hard to copy. So that's my that's my beef with the word strategy. And again, at the end of the day, people just mean plan and that's totally fine. But you want to think a little bit about what can we do that nobody else can.

Pax (27:06)
you you talked about saying like, here's what I want to accomplish and the impact on the business. I think a good place to wrap up would be impact.

it is the giant question mark for people in your industry. I think some good beginnings of how to measure impact of AI search. The fact is, more people are using it than ever, and it will grow over the next coming years. So it'd be foolish for brands to not invest in trying to capture that attention. But I think you see a lot of brands nervous to invest given that they can't see the impact or connect the dots.

completely. And unfortunately, I think you'll see a lot of brands run then to paid because they feel like they can see the dots connect a lot more. How would you recommend, at least today, which we're in July 2026, how would you see brands measure the impact of AI search efforts?

Kevin Indig (28:02)
Okay, here's my riff. So one big transition in this whole AI story is that the old click funnel does not hold anymore, right? So click old click funnel is basically people search, click, convert. That doesn't hold because study after study, including my own, show that people barely click citations, they barely click links. The only exception is AI overviews, where people still see the classic search results, they click a lot more there.

But we're moving towards a world that looks a lot more like ChatGPT than classic Google. And so in that world, people barely click. And so traffic is not the right way to measure success here. And at the same time, it breaks click attribution, right? So the attribution models that we've been using for a long time do not work, at least in organic AI search. So the only attribution that works is self-reported attribution, which is essentially when you ask people after a purchase or a sign-up.

How have you heard about us? And I would argue we can make that a lot more elaborate, we can make it easier, we can make it better, et cetera. But that is the most effective way to truly understand how AI has impacted your business. It's not waterproof, don't get me wrong. People forget, and people, you know, there's all sorts of issues with this. But it's the best way because whenever I do this with my clients, we see values of like up to 10% and sometimes more of net new customers coming through.

AI. So then you see, wait, this is really, really big. Like this is maybe bigger than SEO for for some companies, maybe bigger than organic traffic. There's something to be said about you know monitoring your position in all of that is nice and good. But at the end of the day, self-report attribution.

Is that we have today to measure impact on business.

Pax (29:44)
Love that. thank you, Kevin, so much for for joining us today and and sharing and being generous with your time. we're gonna wrap up with the final question, which we always end with, which is who has had the biggest impact on the way you think about marketing?

Kevin Indig (29:58)
Talk to Dan Petrovic. I mentioned him earlier. I think Dan is a fascinating person, very first principles driven, very experimentation driven. he really does immerse himself as deeply as he can into worlds. And so I think he would be a fascinating person to talk to.

Pax (30:16)
Okay. All right. Yeah. Go please go check him out. Kevin, thank you again for joining. I appreciate the work that you do time you take to further the industry with your studies and research and and being so generous with sharing that information. it's been a pleasure talking with you today.

Kevin Indig (30:31)
Paxton, it's always nice talking to you as well. Looking forward to the next time.

Pax (30:35)
Yep, we'd love to have you back.

Pax (30:37)
That's it for today, everybody. If you enjoyed this episode, please consider leaving us a five-star rating and subscribe so you don't miss future episodes. Big thank you to Kevin Indig for joining us today. You can find Kevin on LinkedIn. Also consider subscribing to his weekly newsletter at growthmemo.com. That's growth-memo.com. You can find past episodes of the campaign and examples of our work at ninety-seventhfloor.com.

There you can learn more about the agency and get in touch with a marketing specialist to get support for your own marketing campaigns. That's it for now. Thank you for listening. As always, keep innovating, keep converting.